v0.24.3

• The README describes what drift does today (#105). It opens with what the package does (fetch land cover for any area, such as a floodplain, compare two dates and every year of a series, turn the change into maps, tables and polygons), shows a map of 2017 to 2023 change on the bundled Neexdzii Kwa reach and a bundled-data example with its real output, and lists all four articles. The DESCRIPTION Description matches it. The docs URL is now https://www.newgraphenvironment.com/drift/ everywhere, and the function reference is grouped by purpose.

v0.24.2

• New article: When Did the Floodplain Change? (#108). It dates the change drift detects on the Bulkley floodplain with dft_rast_break_class(), then looks at each change area in Esri Wayback (sharp images to 2020) and Sentinel-2 composites (after), and says which areas changed on the ground and which are class errors. The article is pkgdown-only; its data comes from data-raw/floodplain_change_dating.R.

v0.24.1

• gdalcubes now comes from NewGraphEnvironment/gdalcubes@newgraph (#104). That build is upstream 0.7.5 plus the fixes proposed as appelmar/gdalcubes#113-115, so a read that fails after the image opens (including a failed cloud-mask read) and a worker that dies early are caught (#99). gdalcubes is a suggested package, so installing drift neither brings it nor replaces an upstream copy: run pak::pak("NewGraphEnvironment/gdalcubes@newgraph"). Remotes: and the install hints revert to appelmar/gdalcubes when those three merge upstream.
  • One #99 case is still not recorded on this build: a worker chunk file the main process cannot read back. inst/notes/gdalcubes-pc-gotchas.md says so.

v0.24.0

• Esri Wayback imagery at the capture nearest a date (#101). Four new functions find, per point or patch, which capture each Wayback release actually shows; a release date is not a capture date. They pick the capture nearest a target date and fetch it as a local cached COG, which QGIS can draw where the Wayback service cannot. Promoted from floodplains#115; on its NECR sample they reproduce every pick (960 of 960) and every re-fetched chip.
  • dft_wayback_releases() lists the releases, sorted by date.
  • dft_wayback_captures(aoi, zoom = 17:19) gives one row per (feature, release, capture). A polygon's cover is the share where that capture is drawn on top.
  • dft_wayback_pick(captures, date) picks the nearest capture. dft_wayback_fetch(aoi, release, zoom) warps it to a nodata-masked RGB COG, and aborts on empty (404) imagery or any GDAL warning.
  • Measured GDAL behaviour shaped the fetch: zoom stops at 22 (GDAL cannot open 23), GDAL's own tile cache is not used (on GDAL 3.13.0 it turns each 404 tile into read errors), and -ovr NONE is passed (otherwise a coarse resolution reads a shallower zoom). Footprints parse with st_make_valid(geos_method = "valid_linework"), because sf's default drops touching parts of multi-part footprints. See research/esri-wayback.md.
  • New Imports: curl, jsonlite. sf is now >= 1.0-8.

v0.23.0

• Cloud-mask failures can no longer put unmasked scenes into a cube or composite without a word (#99).
  • STAC items with no mask asset are dropped. gdalcubes reads a scene whose SCL asset is missing unmasked and records it as fine, so cloud reached the median. dft_stac_cube() and dft_stac_composite() now drop such items before the read and say how many and which. If none has the asset, the read stops with drift_no_items, under its own subclass drift_no_mask_items, and a composite skips that year with that reason rather than "no scenes".
  • Failed reads after an image opens are fixed in gdalcubes, not yet released. #87's check sees only an image that fails to open. A read that fails after opening (a throttled or refused range request), including a corrupted SCL read that leaves the scene unmasked, is still recorded as OK by gdalcubes 0.7.5. The fix is on NewGraphEnvironment/gdalcubes#2 and proposed upstream. drift picks it up unchanged once installed, and its tests for those cases skip until then.
  • A deleted SCL file was already caught. In drift's reads gdalcubes also reads SCL as a band, so a missing SCL fails as an image that would not open and aborts under #87. A test now pins that.
  • The drift_incomplete_cube message now names read failures and unreadable worker files as causes, not only an image that would not open.